Seroter's Daily Reading — #878 (September 30, 2026) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/920285684a21764c1413b4b2f0bf0965e47f35adeddb9be5af670f8e0971865d.mp3

Source: https://seroter.com/2026/09/30/daily-reading-list-september-30-2026-878/

Episode 878 of Seroter's Daily Reading for September 30, 2026. Seroter opens with a personal note: he spent last night at the San Diego Padres playoff game against the Chicago Cubs, which he calls the rowdiest, most fun baseball game he's attended, leaving him hoarse just in time for two major speaking gigs. Then the day's links turn to AI announcements, developer tools, market data, security trends, and voice agents.

First up is Gemini 4 Argon: our next era of frontier intelligence, Google's announcement of its new frontier model. Seroter's take is short and a bit playful: he asks, are you not entertained, notes that Google shared information about the model, and says the numbers look great. The post, by Google DeepMind's Koray Kavukcuoglu, describes Argon as built for long-horizon workflows across software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense. It rolls out first to trusted cyber defenders through the Fairwind Program, with an industry-leading one million token output limit. Google reports internal wins like a forty percent improvement in quantum algorithm optimization, hundreds of terabytes of memory savings across data centers, and large-scale codebase migrations from C and C++ to Rust. Argon also scores seventy-seven point nine percent on DeepSWE, tops the Vals Index for economic impact, and ties for first on CWE-bench for vulnerability remediation. The safety section emphasizes phased release, testing, and safeguards against misuse and prompt injection.

Next is The best AI leaders are ‘a little bit off the wall,’ says Google Cloud executive, a ZDNET piece by Joe McKendrick covering a talk Seroter gave at HubSpot's Unbound conference in Boston. Seroter notes that he did say that, and that the article summarizes some of the points he made. The article lays out his argument that there is no blueprint for AI; everyone is learning as they go. Seroter describes chaotic innovation inside Google, including productivity addiction among top talent who work long days because they are hooked on solving one more problem. He says knowledge work has flipped from twenty percent thinking and eighty percent execution to the reverse, which is mentally exhausting. The best AI leaders, he argues, are a little bit off the wall: not timid, willing to fail constantly for weeks, and protective of the people doing the craziest work. He also compares AI to a crazy intern that you have to steer with real domain knowledge.

Then Build an agentic software factory, starting with one bug. Seroter's commentary is direct: as you build up a set of AI agents to help with software, do it incrementally, and this piece offers a useful lesson. The title itself captures the recommendation: don't start by designing a whole factory; begin with a single bug and let your agentic practice grow from there. This approach reduces risk and keeps the system understandable.

Graph Workflows in ADK: Everything You Need to Know comes from Google Cloud's Annie Wang and Shangjie Chen. Seroter says he liked this explanation of working our way up to more sophisticated agent architectures. The post uses a refund workflow to show how to move from one agent handling everything to a structured graph. It introduces fan-out and fan-in for running independent lookups in parallel, deterministic and agent routers for decisions, human-in-the-loop pauses for manual review, and parallel workers for processing batches of cases. It also covers dynamic orchestration, where Python code schedules follow-up work as results arrive. The key question is whether you can draw the workflow before input arrives; if so, a static edge list is easier to inspect, and if not, dynamic orchestration earns its complexity.

Vercel's State of agent skills offers a data-heavy look at the skills ecosystem. Seroter points out that Vercel sits on a lot of data because it runs the skills.sh registry, and recommends reading this for what skills people are installing, for which industries, and how often. The report says the registry reached one million skills and nearly two hundred eighty million installs in seven months, faster than GitHub, the App Store, or npm reached comparable milestones. Supply leans technical, with more than half of listings for software engineering, agent workflows, data, infrastructure, or security. Demand is more distributed; software engineering still leads installs at eighteen percent, but business operations, writing, and cloud infrastructure get the most installs per listing. The top three hundred seventy-five skills account for sixty-two percent of installs, and seven in eight installs go to cross-industry skills. Vercel argues that the next million skills will teach agents what only your company knows.

State of Markets II is a16z's second chart-packed market report, published by David George. Seroter notes we are getting a bunch of state of such-and-such things lately, and says the a16z folks call out data about companies using AI and more. The report pushes back on a few narratives. It calls tech the everything cycle, noting that tech contributes about seventy-six percent of S&P 500 earnings growth in 2026. It argues that reports of GPU obsolescence are exaggerated: demand for compute still outpaces supply, and even older chips like the A100 are holding value. AI adoption, however, remains shallow: nearly thirty percent of S&P companies report some quantifiable impact, but only about two percent report a tracked metric, and only around two percent of U.S. households pay for an AI service. On software, the report says not apocalypse but prove it: more SaaS companies are profitable, but fewer are growing fast, so multiples compressed. Looking ahead, a16z expects AI to expand the surface area of demand.

Google Threat Intelligence Group's Vulnerability Discovery and Exploitation Trends in the AI Era is the sobering one. Seroter calls it sobering data about the rise in vulnerabilities and the corresponding increase in exploitation, and notes at least there are suggestions for what to do next. The analysis finds that vulnerability disclosures doubled from about five thousand in January 2026 to nearly ten and a half thousand in July and August. Exploitation nearly doubled, from an average of ten point five per month in 2025 to eighteen per month in 2026. Zero-day exploitation is up only marginally, but high-risk vulnerabilities are being weaponized faster. AI-assisted discovery is finding proportionally fewer low-risk issues and more medium and high-risk ones, with half leading to remote code execution. The report also tracks more than two thousand vulnerabilities across the AI stack, with agent orchestration frameworks as the biggest chokepoint. The recommendation is to move from mass patching to threat-intelligence-driven triage, targeted edge defense, and automated agentic remediation.

Voice Agents Can Just Do Things is an essay by Charlie Guo from OpenAI's developer experience team. Seroter says he buys this more than he did a year ago, but he adds a caveat: he doesn't want voice to be the only interface, because sometimes typing is better, or he's in a space where he doesn't want to bark commands at a robot. Guo argues that voice agents do not have to talk back. He describes three modes: speech-to-speech, speech-to-action, and event-to-speech. The underexplored one is speech-to-action, where a user talks and the model uses tools to fill forms, drive creative tools, or operate a computer. Event-to-speech is about proactive, hands-free prompts, like a recipe app telling you when to turn down the heat. Guo also makes the case that native audio models preserve tone, emotion, and cadence that transcription loses. His advice for developers is to expose existing APIs and hooks as tools, and to start by asking what role voice should play.

Finally, Data Agent Kit is now GA: Bring Google Data Cloud to any coding agent. Seroter's summary: do fairly sophisticated analytics and data science work in whatever AI coding tool you prefer. The announcement from Google Cloud explains that Data Agent Kit is a free set of Model Context Protocol tools and agent skills that lets coding agents like Antigravity, Claude Code, or Codex work directly with Google data products. It now supports BigQuery Graph, Bigtable, and Managed Service for Apache Spark on open Lakehouse data. The kit includes more than fifteen Google Data Cloud services, with skills for BigQuery SQL optimization, Bigtable schema design, dbt pipelines, and more. It works across VS Code, Cursor, Cloud Shell, and other environments. The agent can search trusted tables, run queries, design graphs, create schemas, and schedule pipelines using your permissions and governance controls. Seroter's framing emphasizes that this brings sophisticated analytics to whatever coding tool you already use.

Across the day, there's a clear thread: AI models are getting more capable, and the tooling around them is maturing fast, but the hardest questions are still about leadership, judgment, and staying deliberate about how much we hand to the robots.

Your signing session ended